acceptodds
Under review as a conference paper at ICLR 2027

Same Mask, Conflicting Contexts: Learning to Localize AI Edits

Abstract

Knowing that an image has been edited does not specify where it was edited. We study how this distinction should shape the adaptation of spatial visual encoders. Opposite-status context training (OSCT) keeps the recipient's mask fixed while changing the image that supplies attention: in an auxiliary forward, edited recipients read keys and values from unedited donors, and vice versa, with donors drawn from different parent images. Natural and altered-context forwards jointly train the same PE-Spatial localizer; deployment remains an ordinary single-image forward. On a 37,162-image modern-edit evaluation, five 384-update continuations from one shared supervised checkpoint improve edited-image macro F1 by 1.32 points and pixel AP by 1.31 points over matched ordinary training, while reducing pristine-image false-positive area from 0.54% to 0.41%. OSCT also improves F1, IoU, and AP on three external localization sets. Parent-matched donor comparisons support opposite-status selection over same-status exchange, and the improvement persists after matched score refitting of both trained endpoints. These results identify the relationship between borrowed context and recipient supervision as a useful training choice for AI-edit localization, without adding an inference module.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI edit localization, image forensics, dense prediction, attention routing, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Z7pKOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a pipeline that locates the AI edit region in a single image. Specifically, it proposes a new training recipe, which additionally uses an unedited image to route the attention of an edited image from a different source. Results show that the proposed method has a slight gain of F1 score and IoU over the base model PE, while achieves a lower FP rate.

Critical strengths

The paper proposes an interesting idea that uses an unedited image to route the attention of an edited image from a different source. This idea sounds counterintuitive, but experiments show that it improves over the base model and achieves an lower FP rate.

Critical weaknesses
  1. From the reported results, the ordinary PE seems to have a very good base recall and overall F1. The gain achieved by the proposed method is mainly about precision, and is overall marginal considering the slight F1 and IoU improvement.
  2. The method proposed a counterintuitive training recipe. However, there is no in-depth analysis about why it can work. A paper with good analysis can be still very insightful even when the quantitative gain is not high.
Questions for authors
  1. What's the criteria to group two distinct parents together? Do they share similar edits despite being very different images?
  2. What types of edit are present in the training dataset? Common editing types are attribute modification, addition, removal, etc. And can the model generalize to unseen edit type during inference?
  3. How is the real FP pixel area calculated? I assume it's not normalized to [0, 1], but it also doesn't look like the absolute FP pixel counts according to the visual results in Fig.3.
  4. What's the specific settings of PE initialization and ordinary PE?
  5. What does the naming of "modern" and "earlier" mean? Are the images from the modern group being edited by modern AI editor like GPT and QWen, while the editors from the earlier group are not the "modern" models?
  6. As I understand, the proposed auxiliary attention is only with the data from the modern group, and also only present within a certain number of attention blocks. Is this understanding correct? Are there ablations about 1) when training with modern group only / earlier group only / mixed, 2) which group to use the auxiliary attention and 3) in which attention blocks to trigger the proposed fusion?
  7. Does recipient mean the image which accepts the attention key and values from the donor image? Is the recipient-donor definition related to the one in the image forensics literature?
Reviewer_KQPSOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes opposite-status context training (OSCT) for PE-Spatial-G14-448, a transformer-based AI-edit localization model. During an auxiliary training forward, the model routes key and value representations from a different-parent image with the opposite edit status to the recipient, while retaining the recipient’s mask as the target. The authors report that OSCT improves localization metrics over matched training baselines; however, whether this improvement results from reduced reliance on context is not directly established.

Critical strengths

The idea is simple and easy to implement.

The experiments show that the proposed method improves the localization accuracy.

Critical weaknesses

The technical contribution appears limited. Routing context from a different image with the opposite edit status is an interesting idea, but the paper provides limited theoretical analysis of why this strategy should work and offers few additional technical contributions. Moreover, OSCT is implemented and evaluated only with PE-Spatial-G14-448. It is therefore unclear whether the method generalizes to other localization models or encoder architectures.

The motivation of the proposed idea, which is described in the second paragraph of the Introduction, is not sufficiently developed. It is unclear why reliance on image-level edit-status cues would itself be problematic: such cues may be useful for localization, and the paper does not explain why they would lead to inaccurate masks. Likewise, the fact that ordinary mask supervision does not specify which features the model uses is not, by itself, a problem if the model predicts the correct mask. The authors also provide no evidence that the baseline localizer actually relies on global status cues. More fundamentally, the paragraph does not explain how these observations lead to the question of whether the same mask can be predicted under a "borrowed context." The relevance of borrowed context is introduced abruptly, without explaining why context should be varied, how it relates to the hypothesized failure mode, or why training across such contexts would address it. The logical connection from the stated concerns to the proposed question and intervention therefore remains unclear.

The experiments do not sufficiently support the central claim that reducing reliance on context improves localization. Since OSCT changes the training signal, it may act as a mere regularizer and mitigate overfitting in the adapted encoder, including any overfitting associated with model depth. The paper does not distinguish this possibility from the claim that OSCT improves localization by reducing reliance on context.

In addition, the paper uses "PE" and "PE-Spatial" without defining "PE" at first use. Although "PE-Spatial" is cited, the paper also uses "PE" on its own without explanation, making it unclear whether "PE" refers to a model, a model family, or something else. Please define "PE" and clarify its relationship to "PE-Spatial" when first introducing the model.

Questions for authors

Please address the concerns mentioned above.

Reviewer_kXxAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OSCT for AI-edited image localization. During training, the model uses queries from the target image but keys/values from another image with the opposite edit status, while keeping the original segmentation mask unchanged. Inference still uses normal self-attention.

Experiments based on PE-Spatial show consistent improvements over ordinary training on the modern-edit benchmark and three external datasets. The authors also compare opposite-status and same-status donors to study the effect of donor selection.

Critical strengths
  1. The method is simple and doesn't require any additional inference modules. I find the idea of changing the attention context while keeping the segmentation target fixed quite interesting, especially for preventing the model from relying too much on global edit cues.

  2. The experiments are reasonably convincing. The authors use five seeds and matched training settings. Table 3 is particularly useful since same-status exchange does not improve over ordinary training, while OSCT consistently performs better. The improvements on external datasets and the score-refitting results also strengthen the findings.

Critical weaknesses
  1. The improvement is relatively small. On the main benchmark, F1 only increases from 41.26 to 42.58, and AP from 52.17 to 53.48. The results are consistent, but I'm not sure whether this is sufficient for a standalone contribution, especially since the method is only evaluated on one PE-Spatial backbone.

  2. I'm not fully convinced by the explanation of why opposite-status exchange helps. The authors suggest it reduces reliance on image-level status cues, but I didn't see direct evidence for this. The benefit might also come from a stronger regularization effect caused by mixing incompatible features. The same-status control is helpful, but doesn't completely distinguish these explanations.

Questions for authors

Does OSCT also work on other visual backbones, such as DINOv3 or a smaller ViT? I'm curious whether the improvement depends on PE-Spatial's pretrained features.

Have the authors tried random K/V perturbation or attention dropout as additional controls? This might help distinguish the proposed status-conflict mechanism from general regularization.

Can the authors provide per-editor results and image-level false-positive rates on pristine images? The average false-positive pixel area is useful, but doesn't fully describe how often clean images are incorrectly flagged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.